What makes machine learning in healthcare so powerful?

A revolution in healthcare is coming, and it is going to fundamentally change the way we practice and think about medicine.

Ask yourself — what if before you even saw a physician your medical history, blood test results, presenting symptoms, medications over the last year and thousands of other data points had been processed to give a list of potential diagnoses and a recommended course of treatment? Your visit to the doctor could be more personalised, faster, easier, accurate and more focused on your needs.

That is just one of the promises of machine learning (ML) – a subset of artificial intelligence which no longer exists in the realm of science fiction, but is with us right now.

You don’t have to look far for examples – Google Deepmind, IBM Watson and Babylon Health are just a few of the commercial players in the field at the moment, not to mention the countless others leveraging machine learning for business and research (more examples at the end).

So how do they work?


With ML, you give the machine the data and ask it to learn the rules. For example, you could say – “I’m going to show you a bunch of people who had heart attacks, and a bunch who didn’t. Now learn how to tell them apart.”

Machine learning works by taking a dataset of examples labeled with correct predictions. Using this, it “learns” the relationships between the data and the predicted output. Once the machine has reviewed a million patients, you can show it information about a patient it’s never seen before and let it predict whether they may be at imminent risk of a heart attack.

Impact on Healthcare

One unsettling aspect about machine learning from a physician’s perspective is that you can’t actually see the logic the computer has used to reach a conclusion. That can be a challenging mental obstacle, but not an unfamiliar one in medical research. For example – think about the discovery of steroids for immunosuppression. It begins with a very pragmatic observation of, “Oh, this thing works,” and then we tend to backfill our understanding of why it works. That will be the model for a lot of ML applications.

I mentioned before that machine learning is with us right now. So why is it not already widely being used? There are currently two main blockers .

First, the way we store patient data is fragmented, difficult to extract and difficult to contextualise. For example, if you want to know when someone died it can be difficult to piece together if they went home or to another hospital that is not part of your health system. It is possible to overcome, but it takes time and potentially needs new processes.

The bigger obstacle however, is finding a way to take our prediction algorithms and use them safely and responsibly in real world applications that don’t endanger patient lives.

Though the technology for applying machine learning to things like diagnostics exists, it is still really a toy example of what the technology can achieve. The most sophisticated ML algorithm can’t look at a sick child and decide whether they need emergency intubation or whether they can be discharged with conservative therapy. Healthcare professionals synthesise huge amounts of information within milliseconds, often just by “eyeballing” a patient.

The Human Body X-ray Anatomy People Human Skeleton

Even in more direct applications like radiology or pathology where ML can be put to use to interpret CT’s and peripheral blood smears, replacing physicians with software is a long way off. Collecting labeled data for just a single application is extremely time-consuming and expensive. Moreover, datasets require a universal standards to be defined, but gold standards in medicine are often ambiguous.

I believe the goal of machine learning in healthcare will be to help with patient management and logistics. Examples like bed management, heating management and theatre utilisation can help free up time for healthcare professionals to do the most important part of their job – communicate with patients.

We don’t need an AI to replace physician’s – rather we need an AI that serves as a clinical aid part of the clinician’s toolbox for treating patients. The opportunity for machine learning in the medical field lies in logistical patient management, predicting readmissions, triaging patients, auto-populating order sets and any process that could require the interpretation of large trustworthy datasets to provide diagnostic assistance.

The biggest thing I want to emphasise is that it will be clinicians, NOT engineers who are going to push forward the innovations in this field. Right now the limitation is data.

The algorithms used in machine learning are all publicly available, work very well, and can be applied with a moderate programming background. Tensorflow, Theano and Touch are just a few of the technologies in use (learn more). What engineers lack is access to large, high-quality datasets to use them with, and that’s where healthcare professionals can step in.

Written by Dr Adrian Raudaschl

Interesting Recent Research

Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs – JAMA Network, 2016


Dr Adrian Raudaschl is a medical doctor turned product manager. While working in the NHS (National Health Service) he created apps and games to help patients learn more about medical conditions. This demonstrated to Adrian the power great tools have to help people in times when they need it most. Currently he works with children, parents and healthcare professionals to create exciting medical apps and games, which both educate and delight users. Dr Raudaschl is a firm believer in the accessibility of medical information for everyone.